Wearable AI anticipates prolonged sitting in women with chronic pelvic pain

An artificial intelligence system capable of forecasting when women with chronic pelvic pain are likely to remain sedentary could support a more personalised form of digital healthcare.

Researchers at the have developed the approach using activity, heart-rate and sleep data collected through Fitbit devices. Rather than merely recording inactivity after it has occurred, the system attempts to predict a sedentary period in advance. The findings appear in the peer-reviewed, open-access journal npj Women’s Health.

The research points towards digital tools that could deliver a prompt to stand up, stretch or take a short walk shortly before an extended period of sitting begins. However, the study evaluated the accuracy and practicality of the forecasting method. It did not test whether acting on the forecasts reduces pain or produces other clinical benefits.

A condition that can restrict movement

Chronic pelvic pain may be associated with conditions including endometriosis, adenomyosis and uterine fibroids. Pain, fatigue and other symptoms can make it difficult for an individual to remain active throughout the day. The World Health Organization estimates that endometriosis affects approximately 10 per cent of women of reproductive age worldwide, equivalent to around 190 million people. Its symptoms can include severe menstrual pain, chronic pelvic pain, heavy bleeding, abdominal bloating, nausea and fatigue.

Such symptoms can affect work, education, relationships and mental health. This also illustrates why generic instructions to “sit less and move more” may be unsuitable for people managing variable or persistent pain. The Mount Sinai project explored whether a wearable device could instead learn an individual’s normal patterns and identify a suitable time to suggest a manageable amount of movement.

“Our goal was to determine whether everyday wearable devices could serve as an early-warning system for prolonged sitting in women living with chronic pelvic pain,” said senior author Ipek Ensari, Assistant Professor of Artificial Intelligence and Human Health at the Icahn School of Medicine.

“Rather than offering generic advice after the fact, we wanted to determine whether we could anticipate these moments and support people with simple, well-timed prompts that fit naturally into their daily lives.”

Predicting activity one hour ahead

The research team analysed data from 134 women with chronic pelvic pain disorders, primarily endometriosis, together with 61 healthy comparison participants. Participants wore Fitbit devices for up to 90 days, producing minute-by-minute information about physical activity, heart rate and sleep. Approximately ten days of data from each participant were used to train a personalised model capable of forecasting activity levels one hour ahead. The researchers then assessed whether these forecasts could identify 15-minute sedentary periods during waking hours.

The model was not diagnosing a medical event or establishing why someone was inactive. It was identifying patterns in wearable data that suggested a period of limited movement was likely to occur. An eventual intervention might use that forecast to recommend a brief walk or another short period of activity, described by the researchers as an “exercise snack”. The appropriateness of such a prompt would still depend on the user’s symptoms, circumstances and preferences.

A notable finding was that the most computationally complex system was not necessarily the most effective. Relatively simple and interpretable models forecasted sedentary periods about as accurately as the more computationally intensive deep-learning approaches evaluated by the research team.

“We were surprised by how well the simplest models performed,” said lead author Jannes Jegminat, a former postdoctoral research fellow at Mount Sinai. “More complex AI is not always better. Lightweight, interpretable models can accurately forecast sedentary behaviour while being practical enough to run directly on a person’s own device, which also helps protect privacy.”

This finding is relevant to the design of wearable healthcare technology. A lightweight model could potentially operate on a telephone or wearable device instead of repeatedly sending sensitive health information to a remote server. On-device processing does not eliminate every privacy or cybersecurity concern. Nevertheless, reducing the transmission of personal data can limit exposure and decrease dependence on cloud computing. It can also reduce the computational resources needed to deliver a prediction. The approach is consistent with Mount Sinai’s stated emphasis on safe, ethical and human-centred healthcare AI.

Wearable datasets are rarely complete. People remove devices, batteries run out and synchronisation may fail. A system intended for daily life must therefore cope with gaps that would be less common in tightly controlled laboratory research. According to the study release, the forecasting models remained robust when some wearable data were missing. This suggests that useful predictions may still be possible under ordinary conditions of use.

It does not mean that the model will always make a correct forecast. Nor does it establish that the system is ready for routine clinical deployment. The results instead provide evidence that real-world wearable data may be sufficiently dependable to support further development.

From tracking to timely intervention

The next step is to incorporate the forecasting approach into a JITAI. Such systems adjust support according to changing information about a user’s behaviour, condition or surroundings. In this case, a prompt could be issued shortly before a predicted sedentary period rather than at an arbitrary time or after hours of inactivity. Better timing could also reduce notification fatigue. Health applications that issue frequent or irrelevant alerts risk being ignored or disabled. A limited number of personalised prompts may feel less intrusive, although user acceptance will need to be examined formally.

The wider health rationale is well established. The World Health Organization’s guidance on physical activity and sedentary behaviour states that adults should limit sedentary time and notes that greater levels of sedentary behaviour are associated with poorer health outcomes.

However, the Mount Sinai study should not be interpreted as evidence that movement prompts already improve chronic pelvic pain. The researchers say that prospective clinical trials will be required to establish whether predictive prompts reduce sedentary time, alleviate symptoms or improve quality of life.

If those trials are successful, the same forecasting framework might eventually be investigated in other chronic conditions in which pain, fatigue or restricted movement contribute to prolonged inactivity. The immediate significance of the study is more measured but still important. It shows that ordinary wearable data can potentially be converted into an advance warning rather than simply a historical record. It also suggests that practical healthcare AI does not always require the largest or most opaque model. Sometimes the more useful system may be the one that learns a person’s routine, protects their data and knows when not to interrupt.